Re: [PATCH v4 21/22] Add support for pmd_faults

["Kirill A. Shutemov" <kirill@shutemov.name>]

On Mon, Dec 23, 2013 at 07:50:31AM -0700, Matthew Wilcox wrote:
> On Mon, Dec 23, 2013 at 03:41:13PM +0200, Kirill A. Shutemov wrote:
> > > +	/* Fall back to PTEs if we're going to COW */
> > > +	if ((flags & FAULT_FLAG_WRITE) && !(vma->vm_flags & VM_SHARED))
> > > +		return VM_FAULT_FALLBACK;
> > 
> > Why?
> 
> If somebody mmaps a file with MAP_PRIVATE and changes a single byte, I
> think we should allocate a single page to hold that change, not a PMD's
> worth of pages.

We try allocate new huge page in the same situation for AnonTHP. I don't
see a reason why not to do the same here. It would be much harder (if
possible) to collapse small page into a huge one later.

> > > +	pgoff = ((address - vma->vm_start) >> PAGE_SHIFT) + vma->vm_pgoff;
> > > +	size = (i_size_read(inode) + PAGE_SIZE - 1) >> PAGE_SHIFT;
> > > +	if (pgoff >= size)
> > > +		return VM_FAULT_SIGBUS;
> > > +	if ((pgoff | PG_PMD_COLOUR) >= size)
> > > +		return VM_FAULT_FALLBACK;
> > 
> > I don't think it's necessary to fallback in this case.
> > Do you care about SIGBUS behaviour or what?
> 
> I'm looking to preserve the same behaviour we see with PTE mappings.  I mean,
> it's supposed to be _transparent_ huge pages, right?

We can't be totally transparent. At least from performance point of view.

The question is whether it's critical to preserve SIGBUS beheviour. I
would prefer to map last page in mapping with huge pages too, if it's
possible.

Do you know anyone who relay on SIGBUS for correctness?

> 
> > > + insert:
> > > +	length = xip_get_pfn(inode, &bh, &pfn);
> > > +	if (length < 0)
> > > +		return VM_FAULT_SIGBUS;
> > > +	if (length < PMD_SIZE)
> > > +		return VM_FAULT_FALLBACK;
> > > +	if (pfn & PG_PMD_COLOUR)
> > > +		return VM_FAULT_FALLBACK;	/* not aligned */
> > 
> > Without assistance from get_unmapped_area() you will hit this all the time
> > (511 of 512 on x86_64).
> 
> Yes ... I thought you were working on that part for your transparent huge
> page cache patchset?

Yeah, I have patch for x86-64. Just a side note.

> 
> > And the check should be moved before get_block(), I think.
> 
> Can't.  The PFN we're checking is the PFN of the storage.  We have to
> call get_block() to find out where it's going to be.

I see.

-- 
 Kirill A. Shutemov

--
To unsubscribe, send a message with 'unsubscribe linux-mm' in
the body to majordomo@kvack.org.  For more info on Linux MM,
see: http://www.linux-mm.org/ .
Don't email: <a href=mailto:"dont@kvack.org"> email@kvack.org </a>